<html xmlns:o="urn:schemas-microsoft-com:office:office" xmlns:w="urn:schemas-microsoft-com:office:word" xmlns:m="http://schemas.microsoft.com/office/2004/12/omml" xmlns="http://www.w3.org/TR/REC-html40"> <head> <meta http-equiv="Content-Type" content="text/html; charset=Windows-1252"> <meta name="Generator" content="Microsoft Word 15 (filtered medium)"> <style></style> </head> <body lang="EN-US" link="blue" vlink="purple" style="word-wrap:break-word"> <div class="WordSection1"> I like the idea of doing just one STUN request to avoid the many semi-duplicate candidates.<o:p></o:p> <o:p> </o:p> Another interesting thing to think about is that in most cases (99.9% ish) the STUN query is going to return the same result as last time. In most cases the host computer has not moved networks and the external address has not changed.<o:p></o:p> <o:p> </o:p> What if libnice could remember the previous external address and lead with that as a candidate. Libnice could still do one (or more) STUN queries to check if the external address has changed. But starting with the previously known external address could speed up connection a lot.<o:p></o:p> <o:p> </o:p> The challenge is “how to remember the previous external address”. Libnice lives in somebody’s process and on some random host computer. Libnice might be completely shut down in-between uses, even if the process keeps running.<o:p></o:p> <o:p> </o:p> What if libnice remembered the previous external address somewhere in process space. If the process shuts down the knowledge is lost. But if the process keeps running (e.g. a server or a long running browser) then the external address is remembered. Is there a cross-platform place to stash data? Environment variables might work. A persistent background thread might work.<o:p></o:p> <o:p> </o:p> The reason I suggest optimizations like this is customer experience. I see chrome to gstreamer establish connections in less than a second. Gstreamer to gstreamer often takes much more time – easily five seconds plus.<o:p></o:p> <o:p> </o:p> Stuart<o:p></o:p> <o:p> </o:p> <div style="border:none;border-top:solid #B5C4DF 1.0pt;padding:3.0pt 0in 0in 0in"> From: nice <nice-bounces@lists.freedesktop.org> Date: Monday, October 19, 2020 at 11:00 AM To: Stuart Marshall <stuart@seelye.net> Cc: Juan Navarro <juan.navarro@gmx.es>, nice@lists.freedesktop.org <nice@lists.freedesktop.org> Subject: Re: [libnice] Force an external address (srflx) candidate?<o:p></o:p> </div> <div> Hi Stuart, On 10/14/20 at 09:22pm, Stuart Marshall wrote: > In contrast, the ICE candidates emitted by Chrome are stunningly few and precise. I understand that the ICE protocols and the libnice implementation were/are meant to be general case. But they miss obvious efficiencies that can be provided by additional external information. STUN servers facilitate some of that additional information, but introduce a dependency and more latency. > > My knowledge of libnice internals is not great, but I kind of wish we could > > 1. Feed some particular IP candidate addresses to it, > 2. Tell it to skip a bunch of other candidate generation and testing I think of some optimisations, that could help to limit the number of candidates, without weakening the versatility of the ice method overall (except in *rare* cases where the server running libnice uses source routing, ie chooses the default route based on the source IP address): We could use a single server reflexive candidate and relay local candidate, per stream/component. Generally, there's no gain to send a stun request from each local interface, because all packets will reach the same stun server by the same default route. The consequence is that we obtain <N> distinct server-reflexive candidates from <N> distinct source IP addresses. These server-reflexives candidates are distinct because their IP address will be the same (this is our public IP address), but the port mapping will be different. The same applies to turn relay candidates too (including unnecessary resources reservations on the turn servers BTW). To avoid that, we could for example: 1. discard these redundant candidates when we discover them (when processing the discovery stun response). 2. or more radically, just send a single stun and turn discovery request. In case of 2. the choice of the local interface used as the base address to send this unique stun/turn discoevry request is normally not relevant, because the routing table will hopefully make these packets go out by the same default route again, whatever source interface they come from. To summarize, I think that sending a single stun request from a single network interface during gathering phase to obtain our server-reflexive address is normally a cheap operation (one RTT when the thr stun server is available), but what is expensive from libnice point of view is to deal with many identical reflexive/relay candidates during the connecting phase, because it creates many possibilities to be tested. And the more possibilities we have to test, the more time it takes to complete. Best wishes, -- fabrice _______________________________________________ nice mailing list nice@lists.freedesktop.org <a href="https://lists.freedesktop.org/mailman/listinfo/nice">https://lists.freedesktop.org/mailman/listinfo/nice</a><o:p></o:p> </div> </div> </body> </html>